class Agent(object):
def __init__(self, name='model', input_num=None, output_num=None):
"""A learning agent that uses tensorflow to create a neural network"""
assert input_num is not None
assert output_num is not None
self.input_num = input_num
self.output_num = output_num
self._build_net()
def _build_net(self):
"""Construct the neural network"""
# Change the network structure here
S = Input(shape=[self.input_num])
h0 = Dense(300, activation="sigmoid")(S)
h1 = Dense(600, activation="sigmoid")(h0)
h2 = Dense(29, activation="sigmoid")(h1)
V = Dense(self.output_num, activation="sigmoid")(h2)
self.model = Model(inputs=S, outputs=V)
self.model.compile(optimizer="adam", loss='mse')
def train(self, x, y, n_epoch=100, batch=32):
"""Train the network"""
self.model.fit(x=x, y=y, epochs=n_epoch, batch_size=batch)
def predict(self, x):
"""Input values to the neural network and return the result"""
a = self.model.predict(x)
return a
def test_add_entropy_loss_on_functional_model(self):
inputs = Input(shape=(1, ))
targets = Input(shape=(1, ))
outputs = testing_utils.Bias()(inputs)
model = Model([inputs, targets], outputs)
model.add_loss(losses.binary_crossentropy(targets, outputs))
model.compile('sgd', run_eagerly=testing_utils.should_run_eagerly())
with test.mock.patch.object(logging, 'warning') as mock_log:
model.fit([self.x, self.y], batch_size=3, epochs=5)
self.assertNotIn('Gradients do not exist for variables',
str(mock_log.call_args))
class NN:
"""
The NN class wraps a keras Sequential model to reduce the interface methods
Notice:
Difference to dqn is just the setter and getter methods for the weights
"""
def __init__(self, env, atoms, alpha: float = 0.001, decay: float = 0.0001):
"""
We initialize our functional model, therefore we need Input Shape and Output Shape
:param env:
:param alpha:
:param decay:
"""
self.alpha = alpha
self.decay = decay
self.model = None
self.atoms = atoms
# new to D-DDQN
self.init_model(env.observation_space.shape[0], env.action_space.n)
def init_model(self, input_shape: int, n_actions: int):
"""
Initializing our keras sequential model
:return: initialized model
"""
input = Input(shape=(input_shape,))
h1 = Dense(64, activation='relu')(input)
h2 = Dense(64, activation='relu')(h1)
outputs = []
for _ in range(n_actions):
outputs.append(Dense(self.atoms, activation='softmax')(h2))
self.model = Model(input, outputs)
def predict(self, *args, **kwargs):
"""
By wrapping the keras predict method we can handle our net as a standalone object
:param args: interface to keras.model.predict
:return: prediction
"""
return self.model.predict(*args, **kwargs)
def fit(self, *args, **kwargs):
"""
By wrapping the keras fit method we can handle our net as a standalone object
:param args: interface to keras.model.fit
:return: history object
"""
return self.model.fit(*args, **kwargs)
def get_weights(self):
"""
Passing the arguments to keras get_weights
"""
return self.model.get_weights()
def set_weights(self, *args, **kwargs):
"""
Passing the arguments to keras set_weights
"""
self.model.set_weights(*args, *kwargs)
def test_loss_with_sample_weight_in_layer_call(self):
class MyLayer(layers.Layer):
def __init__(self):
super(MyLayer, self).__init__()
self.bias = testing_utils.Bias()
def call(self, inputs):
out = self.bias(inputs[0])
self.add_loss(MAE()(inputs[1], out, inputs[2]))
self.add_loss(
math_ops.reduce_mean(inputs[2] * mae(inputs[1], out)))
return out
inputs = Input(shape=(1, ))
targets = Input(shape=(1, ))
sw = Input(shape=(1, ))
outputs = MyLayer()([inputs, targets, sw])
model = Model([inputs, targets, sw], outputs)
model.predict([self.x, self.y, self.w])
model.compile(optimizer_v2.gradient_descent.SGD(0.05),
run_eagerly=testing_utils.should_run_eagerly(),
experimental_run_tf_function=testing_utils.
should_run_tf_function())
history = model.fit([self.x, self.y, self.w], batch_size=3, epochs=5)
self.assertAllClose(history.history['loss'], [2., 1.8, 1.6, 1.4, 1.2],
1e-3)
output = model.evaluate([self.x, self.y, self.w])
self.assertAlmostEqual(output, 1.0, 3)
output = model.test_on_batch([self.x, self.y, self.w])
self.assertAlmostEqual(output, 1.0, 3)
def fronzen():
x = Input(shape=(32, ))
layer = Dense(32)
layer.trainable = False
y = layer(x)
frozen_model = Model(x, y)
# 在下面的模型中，训练期间不会更新层的权重
frozen_model.compile(optimizer='rmsprop', loss='mse')
layer.trainable = True
trainable_model = Model(x, y)
# 使用这个模型，训练期间 `layer` 的权重将被更新
# (这也会影响上面的模型，因为它使用了同一个网络层实例)
trainable_model.compile(optimizer='rmsprop', loss='mse')
frozen_model.fit(data, labels) # 这不会更新 `layer` 的权重
trainable_model.fit(data, labels) # 这会更新 `layer` 的权重
def train_model(model: Model, images):
time_str = datetime.datetime.now().strftime("%y-%m-%d_%H-%M-%S")
callbacks = [
keras.callbacks.TensorBoard(f'{LOGS_DIR}{time_str}'),
TensorBoardImage(f'{LOGS_DIR}{time_str}', "Emojis", images,
period=100),
CheckpointCallback(f'{LOGS_DIR}{time_str}', period=100),
]
model.fit(
images,
images,
epochs=100000,
batch_size=len(images),
# validation_data=(images, images),
callbacks=callbacks,
verbose=0)
model.save(f"../logs/{time_str}/model.h5")
def attn_many_to_one(dataset_object: LSTM_data):
X_train, X_test, Y_train, Y_test = dataset_object.get_memory()
X_train, X_test = X_train[:, :, :-12], X_test[:, :, :-12]
i = Input(shape=(X_train.shape[1], X_train.shape[2]))
att_in = LSTM(NEURONS,
return_sequences=True,
activation=ACTIVATION,
recurrent_activation="sigmoid",
activity_regularizer=regularizers.l2(L2),
bias_regularizer=regularizers.l2(BIAIS_REG),
)(i)
att_in = LSTM(NEURONS,
return_sequences=True,
activation=ACTIVATION,
recurrent_activation="sigmoid",
activity_regularizer=regularizers.l2(L2),
bias_regularizer=regularizers.l2(BIAIS_REG),
)(att_in)
att_in = LSTM(NEURONS,
return_sequences=True,
activation=ACTIVATION,
recurrent_activation="sigmoid",
activity_regularizer=regularizers.l2(L2),
bias_regularizer=regularizers.l2(BIAIS_REG),
)(att_in)
att_out = attention()(att_in)
att_out = Dropout(DROPOUT)(att_out)
outputs = Dense(1,
activation='relu',
trainable=True,
bias_regularizer=regularizers.l2(BIAIS_REG),
activity_regularizer=regularizers.l2(L2)
)(att_out)
model = Model(inputs=[i], outputs=[outputs])
optim = Adam()
model.compile(optimizer=optim,
loss=['mean_squared_error']
)
# Fitting the RNN to the Training set
history = model.fit(X_train, Y_train,
epochs=EPOCHS,
batch_size=BATCH_SIZE,
validation_data=(X_test, Y_test),
callbacks=[EARLY_STOP, REDUCE_LR]
)
model.save("data/weights/attn_based_lstm_no_senti")
plot_train_loss(history)
evaluate(model,X_test,Y_test, dataset_object,name="attn_evaluate", senti="no")
def run_model(model_fn,
optimizer='adam',
loss='binary_crossentropy',
steps_per_epoch=1,
epochs=1):
inputs, outputs = model_fn(128, 128, 1)
_model = Model(inputs=[inputs], outputs=[outputs])
_model.compile(optimizer=optimizer, loss=loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# Write the session graph to the logs.
tf.summary.FileWriter('./logs/', sess.graph)
_model.fit(TensorFeed().build_dataset().dataset,
steps_per_epoch=steps_per_epoch,
epochs=epochs)
return _model
def test_loss_on_model_fit(self):
inputs = Input(shape=(1,))
targets = Input(shape=(1,))
outputs = testing_utils.Bias()(inputs)
model = Model([inputs, targets], outputs)
model.add_loss(MAE()(targets, outputs))
model.add_loss(math_ops.reduce_mean(mae(targets, outputs)))
model.compile(
optimizer_v2.gradient_descent.SGD(0.05),
run_eagerly=testing_utils.should_run_eagerly())
history = model.fit([self.x, self.y], batch_size=3, epochs=5)
self.assertAllClose(history.history['loss'], [2., 1.8, 1.6, 1.4, 1.2], 1e-3)
def test_loss_with_sample_weight_on_model_fit(self):
inputs = Input(shape=(1,))
targets = Input(shape=(1,))
sw = Input(shape=(1,))
outputs = testing_utils.Bias()(inputs)
model = Model([inputs, targets, sw], outputs)
model.add_loss(MAE()(targets, outputs, sw))
model.add_loss(3 * math_ops.reduce_mean(sw * mae(targets, outputs)))
model.compile(
optimizer_v2.gradient_descent.SGD(0.025),
run_eagerly=testing_utils.should_run_eagerly())
history = model.fit([self.x, self.y, self.w], batch_size=3, epochs=5)
self.assertAllClose(history.history['loss'], [4., 3.6, 3.2, 2.8, 2.4], 1e-3)
def train_model(model: Model, epochs: int) -> dict:
model.compile(optimizer='adam',
loss='mean_squared_error',
metrics=[metrics.CategoricalAccuracy(), metrics.Recall(), metrics.Precision()])
y_train = to_categorical(globals()['TRAINING_LABELS'])
y_validate = to_categorical(globals()['VALIDATION_LABELS'])
print(y_train.shape)
history = model.fit(x=globals()['TRAINING_DATA'], y=y_train, batch_size=32, epochs=epochs,
validation_data=(globals()['VALIDATION_DATA'], y_validate),
validation_batch_size=32,
validation_freq=1)
return history.history
def train_autoencoder(x: np.ndarray, cfg: dict, autoencoder: Model) -> History:
r"""Train an already built AE model on new data.
:param x: The data the AE shall be trained on.
:param cfg: ConfigurationSpace values that were used to construct this AE.
:param autoencoder: The constructed AE model.
:return: The training history.
"""
callbacks = None
if cfg['ae_type'] == 'deep_ksparse':
callbacks = [UpdateKSparseLevel()]
return autoencoder.fit(x,
x,
callbacks=callbacks,
epochs=cfg['epochs'],
batch_size=BATCH_SIZE)
class GenericModel:
@staticmethod
def load_from(path):
model = GenericModel()
model.model = load_model(path)
return model
def __init__(self):
self.model = None
self.registered_callbacks = []
self.id = 'generic_model'
self.time = round(time())
self.desc = None
"""config = ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.40
config.gpu_options.allow_growth = True
session = InteractiveSession(config=config)"""
def build_model(self):
img_input = Input(self.get_input_shape())
last_layer = self.model_structure(img_input)
self.model = Model(img_input, last_layer)
self.model.summary()
def compile(self,
loss_function,
metric_functions=None,
optimizer=Adam(1e-3, epsilon=1e-6)):
self.require_model_loaded()
return self.model.compile(loss=loss_function,
optimizer=optimizer,
metrics=metric_functions)
def model_structure(self, input_img):
raise NotImplementedError
def get_input_shape(self):
raise NotImplementedError
def register_std_callbacks(self,
tensorboard_logs_folder=None,
checkpoint_path=None):
self.require_model_loaded()
run_id = str(time())
if self.desc is not None:
run_id += "_" + self.desc
folder_id = os.path.join(self.id, run_id)
if tensorboard_logs_folder is not None:
self.registered_callbacks.append(
TensorBoard(log_dir=os.path.join(tensorboard_logs_folder,
folder_id),
histogram_freq=0,
write_graph=True,
write_images=True))
if checkpoint_path is not None:
store_path = os.path.join(checkpoint_path, folder_id)
if not os.path.exists(store_path):
os.makedirs(store_path)
store_path = os.path.join(
store_path, 'e{epoch:02d}-l{loss:.4f}-v{val_loss:.4f}.ckpt')
print("Storing to %s" % store_path)
self.registered_callbacks.append(
ModelCheckpoint(store_path,
monitor='val_loss',
verbose=1,
period=1,
save_best_only=False,
mode='min'))
def train_with_generator(self,
training_data_generator,
epochs,
steps_per_epoch,
validation_data=None):
self.model.fit(training_data_generator,
use_multiprocessing=True,
workers=4,
steps_per_epoch=steps_per_epoch,
callbacks=self.registered_callbacks,
epochs=epochs,
verbose=1,
**({} if validation_data is None else {
"validation_data": validation_data
}))
def require_model_loaded(self):
if self.model is None:
raise ValueError("Model is not build yet")
def load_weights(self, path):
self.require_model_loaded()
return self.model.load_weights(path)
def predict(self, batch):
self.require_model_loaded()
return self.model.predict(batch)
def train(review_data):
################################################################
# declare input embeddings to the model
#User input
user_id_input = Input(shape=[1], name='user')
#Item Input
item_id_input = Input(shape=[1], name='item')
price_id_input = Input(shape=[1], name='price')
title_id_input = Input(shape=[1], name='title')
# define the size of embeddings as a parameter
# ****H: size_of_embedding - 5, 10 , 15, 20, 50
size_of_embedding = 15
user_embedding_size = size_of_embedding
item_embedding_size = size_of_embedding
price_embedding_size = size_of_embedding
title_embedding_size = size_of_embedding
# apply an embedding layer to all inputs
user_embedding = Embedding(output_dim=user_embedding_size,
input_dim=users.shape[0],
input_length=1,
name='user_embedding')(user_id_input)
item_embedding = Embedding(output_dim=item_embedding_size,
input_dim=items_reviewed.shape[0],
input_length=1,
name='item_embedding')(item_id_input)
price_embedding = Embedding(output_dim=price_embedding_size,
input_dim=price.shape[0],
input_length=1,
name='price_embedding')(price_id_input)
title_embedding = Embedding(output_dim=title_embedding_size,
input_dim=titles.shape[0],
input_length=1,
name='title_embedding')(title_id_input)
# reshape from shape (batch_size, input_length,embedding_size) to (batch_size, embedding_size).
user_vecs = Reshape([user_embedding_size])(user_embedding)
item_vecs = Reshape([item_embedding_size])(item_embedding)
price_vecs = Reshape([price_embedding_size])(price_embedding)
title_vecs = Reshape([title_embedding_size])(title_embedding)
################################################################
# Concatenate the item embeddings :
item_vecs_complete = Concatenate()([item_vecs, price_vecs, title_vecs])
# Concatenate user and item embeddings and use them as features for the neural network:
input_vecs = Concatenate()([user_vecs, item_vecs_complete
]) # can be changed by Multiply
#input_vecs = Concatenate()([user_vecs, item_vecs]) # can be changed by Multiply
# Multiply user and item embeddings and use them as features for the neural network:
#input_vecs = Multiply()([user_vecs, item_vecs]) # can be changed by concat
# Dropout is a technique where randomly selected neurons are ignored during training to prevent overfitting
input_vecs = Dropout(0.1)(input_vecs)
# Check one dense 128 or two dense layers (128,128) or (128,64) or three denses layers (128,64,32))
# First layer
# Dense(128) is a fully-connected layer with 128 hidden units.
# Use rectified linear units (ReLU) f(x)=max(0,x) as an activation function.
x = Dense(128, activation='relu')(input_vecs)
x = Dropout(0.1)(x) # Add droupout or not # To improve the performance
# Next Layers
#x = Dense(128, activation='relu')(x) # Add dense again or not
x = Dense(64, activation='relu')(x) # Add dense again or not
x = Dropout(0.1)(x) # Add droupout or not # To improve the performance
x = Dense(32, activation='relu')(x) # Add dense again or not #
x = Dropout(0.1)(x) # Add droupout or not # To improve the performance
# The output
y = Dense(1)(x)
################################################################
model = Model(
inputs=[user_id_input, item_id_input, price_id_input, title_id_input],
outputs=y)
################################################################
# ****H: loss
# ****H: optimizer
model.compile(loss='mse', optimizer="adam")
################################################################
save_path = "./"
mytime = time.strftime("%Y_%m_%d_%H_%M")
# modname = 'dense_2_15_embeddings_2_epochs' + mytime
modname = 'dense_2_15_embeddings_2_epochs'
thename = save_path + '/' + modname + '.h5'
mcheck = ModelCheckpoint(thename, monitor='val_loss', save_best_only=True)
################################################################
# ****H: batch_size
# ****H: epochs
# ****H:
# ****H:
history = model.fit([
ratings_train["user_id"], ratings_train["item_id"],
ratings_train["price_id"], ratings_train["title_id"]
],
ratings_train["score"],
batch_size=64,
epochs=2,
validation_split=0.2,
callbacks=[mcheck],
shuffle=True)
print("MSE: ", history.history)
return model
class jyHEDModelV1(jyModelBase):
def __init__(self):
super(jyHEDModelV1, self).__init__()
self.__listLayerName = []
self.__pVisualModel = None
def structureModel(self):
Inputs = layers.Input(shape=self._inputShape, batch_size=self._iBatchSize)
Con1 = layers.Conv2D(64, (3, 3), name='Con1', activation='relu', padding='SAME', input_shape=self._inputShape, strides=1)(Inputs)
Con2 = layers.Conv2D(64, (3, 3), name='Con2', activation='relu', padding='SAME', strides=1)(Con1)
Side1 = sideBranch(Con2, 1)
MaxPooling1 = layers.MaxPooling2D((2, 2), name='MaxPooling1', strides=2, padding='SAME')(Con2)
# outputs1
Con3 = layers.Conv2D(128, (3, 3), name='Con3', activation='relu', padding='SAME', strides=1)(MaxPooling1)
Con4 = layers.Conv2D(128, (3, 3), name='Con4', activation='relu', padding='SAME', strides=1)(Con3)
Side2 = sideBranch(Con4, 2)
MaxPooling2 = layers.MaxPooling2D((2, 2), name='MaxPooling2', strides=2, padding='SAME')(Con4)
# outputs2
Con5 = layers.Conv2D(256, (3, 3), name='Con5', activation='relu', padding='SAME', strides=1)(MaxPooling2)
Con6 = layers.Conv2D(256, (3, 3), name='Con6', activation='relu', padding='SAME', strides=1)(Con5)
Con7 = layers.Conv2D(256, (3, 3), name='Con7', activation='relu', padding='SAME', strides=1)(Con6)
Side3 = sideBranch(Con7, 4)
MaxPooling3 = layers.MaxPooling2D((2, 2), name='MaxPooling3', strides=2, padding='SAME')(Con7)
# outputs3
Con8 = layers.Conv2D(512, (3, 3), name='Con8', activation='relu', padding='SAME', strides=1)(MaxPooling3)
Con9 = layers.Conv2D(512, (3, 3), name='Con9', activation='relu', padding='SAME', strides=1)(Con8)
Con10 = layers.Conv2D(512, (3, 3), name='Con10', activation='relu', padding='SAME', strides=1)(Con9)
Side4 = sideBranch(Con10, 8)
MaxPooling4 = layers.MaxPooling2D((2, 2), name='MaxPooling4', strides=2, padding='SAME')(Con10)
# outputs4
Con11 = layers.Conv2D(512, (3, 3), name='Con11', activation='relu', padding='SAME', strides=1)(MaxPooling4)
Con12 = layers.Conv2D(512, (3, 3), name='Con12', activation='relu', padding='SAME', strides=1)(Con11)
Con13 = layers.Conv2D(512, (3, 3), name='Con13', activation='relu', padding='SAME', strides=1)(Con12)
Side5 = sideBranch(Con13, 16)
Fuse = layers.Concatenate(axis=-1)([Side1, Side2, Side3, Side4, Side5])
# learn fusion weight
Fuse = layers.Conv2D(1, (1, 1), name='Fuse', padding='SAME', use_bias=False, activation=None)(Fuse)
output1 = layers.Activation('sigmoid', name='output1')(Side1)
output2 = layers.Activation('sigmoid', name='output2')(Side2)
output3 = layers.Activation('sigmoid', name='output3')(Side3)
output4 = layers.Activation('sigmoid', name='output4')(Side4)
output5 = layers.Activation('sigmoid', name='output5')(Side5)
output6 = layers.Activation('sigmoid', name='output6')(Fuse)
outputs = [output1, output2, output3, output4, output5, output6]
self._pModel = Model(inputs=Inputs, outputs=outputs)
pAdam = optimizers.adam(lr=0.0001)
self._pModel.compile(loss={
'output6': classBalancedSigmoidCrossEntropy
}, optimizer=pAdam)
# self._pModel.summary()
def startTrain(self, listDS, iMaxLen, iBatchSize):
itrTrain = tf.compat.v1.data.make_one_shot_iterator(listDS[0])
itrValid = tf.compat.v1.data.make_one_shot_iterator(listDS[1])
iStepsPerEpochTrain = int(iMaxLen[0] / iBatchSize[0])
iStepsPerEpochValid = int(iMaxLen[1] / iBatchSize[1])
self._pModel.fit(itrTrain, validation_data=itrValid, epochs=self._iEpochs,
callbacks=[self._pSaveModel, self._pTensorboard], steps_per_epoch=iStepsPerEpochTrain,
validation_steps=iStepsPerEpochValid)
def loadWeights(self, strPath):
# last = tf.train.latest_checkpoint(strPath)
# checkPoint = tf.train.load_checkpoint(strPath)
self._pModel.load_weights(strPath)
# visual model
outputs = []
for myLayer in self._pModel.layers:
self.__listLayerName.append(myLayer.name)
outputs.append(myLayer.output)
# print(self.__pModel.layers[0])
# self.__pVisualModel = Model(self.__pModel.inputs, outputs=outputs)
self.__pVisualModel = Model(self._pModel.inputs, outputs=self._pModel.outputs)
return self.__pVisualModel
def predict(self, IMG):
# pImage = open(IMG, 'rb').read()
# tensorIMG = tf.image.decode_jpeg(pImage)
pIMG = image.array_to_img(IMG)# .resize((256, 144))
tensorIMG = image.img_to_array(pIMG)
x = np.array(tensorIMG / 255.0)
# show image
iColumn = 4
# generate window
plt.figure(num='Input')
# plt.subplot(1, 1, 1)
plt.imshow(x)
# imagetest = x
x = np.expand_dims(x, axis=0)
# pyplot.imshow(x)
time1 = datetime.datetime.now()
outputs = self.__pVisualModel.predict(x)
time2 = datetime.datetime.now()
print(time2 - time1)
i = 100
listOutput = []
for i in range(len(outputs)):
outputShape = outputs[i].shape
singleOut = outputs[i].reshape(outputShape[1], outputShape[2], outputShape[3])
# singleOut *= 255
listOutput.append(singleOut)
singleOut = listOutput[-1]
singleOut[singleOut > 0.5] = 1
listOutput[-1] = singleOut
return listOutput
'''
for output in outputs:
# plt.figure(num='%s' % str(i))
outputShape = output.shape
singleOut = output.reshape(outputShape[1], outputShape[2], outputShape[3])
singleOut *= 255
if outputShape[3] == 1:
# test = x - output
# test = np.abs(test)
# return mysum
# plt.subplot(1, 1, 1)
# plt.imshow(singleOut, camp='gray')
# cv2.imwrite('D:\wyc\Projects\TrainDataSet\HED\Result/%s.jpg' % str(i), singleOut)
return singleOut
# i += 1
# plt.show()
'''
def getModelConfig(self):
return self._iBatchSize
class PPO:
def __init__(self,
n_actions,
n_features,
actor_lr=0.0001,
critic_lr=0.0001,
reward_decay=0.9,
l2=0.001,
loss_clipping=0.2,
target_update_alpha=0.9):
self.n_actions = n_actions
self.n_features = n_features
self.actor_lr = actor_lr # 学习率
self.critic_lr = critic_lr
self.gamma = reward_decay # reward 递减率
self.states, self.actions, self.rewards, self.states_, self.dones, self.v_by_trace = [], [], [], [], [], [] # V(s)=r+g*V(s_)
self.l2 = l2
self.loss_clipping = loss_clipping
self.target_update_alpha = target_update_alpha # 模型参数平滑因子
self._build_critic()
self.actor = self._build_actor()
self.actor_old = self._build_actor()
self.actor_old.set_weights(self.actor.get_weights())
self.dummy_advantage = np.zeros((1, 1))
self.dummy_old_prediction = np.zeros((1, self.n_actions))
def _build_critic(self):
inputs = Input(shape=(self.n_features, ))
x = Dense(32, activation='relu',
kernel_regularizer=l2(self.l2))(inputs)
x = Dense(16, activation='relu', kernel_regularizer=l2(self.l2))(x)
output = Dense(1, kernel_regularizer=l2(self.l2))(x)
self.critic = Model(inputs=inputs, outputs=output)
self.critic.compile(optimizer=Adam(lr=self.critic_lr),
loss='mean_squared_error',
metrics=['accuracy'])
def _build_actor(self):
state = Input(shape=(self.n_features, ), name="state")
advantage = Input(shape=(1, ), name="Advantage")
old_prediction = Input(shape=(self.n_actions, ), name="Old_Prediction")
x = Dense(32, activation='relu', kernel_regularizer=l2(self.l2))(state)
x = Dense(16, activation='relu', kernel_regularizer=l2(self.l2))(x)
policy = Dense(self.n_actions,
activation='softmax',
kernel_regularizer=l2(self.l2))(x)
model = Model(inputs=[state, advantage, old_prediction],
outputs=policy)
model.compile(optimizer=Adam(lr=self.actor_lr),
loss=self.proximal_policy_optimization_loss(
advantage=advantage, old_prediction=old_prediction))
return model
def proximal_policy_optimization_loss(self, advantage, old_prediction):
def loss(y_true, y_pred):
prob = y_true * y_pred
old_prob = y_true * old_prediction
r = prob / (old_prob + 1e-10)
return -K.mean(
K.minimum(
r * advantage,
K.clip(r,
min_value=1 - self.loss_clipping,
max_value=1 + self.loss_clipping) * advantage))
# + 0.2 * (prob * K.log(prob + 1e-10)))
return loss
def choose_action(self, observation, is_train_mode=True):
observation = np.array(observation)
observation = observation[np.newaxis, :]
action_probs = self.actor.predict(
[observation, self.dummy_advantage, self.dummy_old_prediction])
# print('action_probs', action_probs)
if is_train_mode:
action = int(
np.random.choice(range(action_probs.shape[1]),
p=np.squeeze(action_probs))) # 加入了随机性
else:
action = int(np.squeeze(np.argmax(action_probs, axis=1)))
return action
def store_transition(self, s, a, r, s_, d):
self.states.append(s)
self.actions.append(a)
self.rewards.append(r)
self.states_.append(s_)
self.dones.append(d)
def learn(self):
# print('learn: sample length-',len(self.actions))
# print('learn: states-',self.states)
# print('learn: actions-',self.actions)
self.cal_v_by_traceback()
b_s, b_a, b_vt = np.array(self.states), np.array(
self.actions), np.array(self.v_by_trace)
b_v = self.get_v(b_s)
# print('b_s:{}'.format(self.states))
# print('b_a:{}'.format(self.actions))
# print('b_r:{}'.format(self.rewards))
# print('b_d:{}'.format(self.dones))
# print('b_vt:{}'.format(self.v_by_trace))
# print('b_v:{}'.format(b_v))
b_adv = b_vt - b_v # 可以对adv做标准化
# b_adv = (b_adv - np.mean(b_adv)) / (np.std(b_adv) + 1e-10)
b_old_prediction = self.get_old_prediction(b_s)
b_a_onehot = np.zeros((b_a.shape[0], self.n_actions))
b_a_onehot[:, b_a.flatten()] = 1
# print('b_adv:{}'.format(b_adv))
# print('b_old_prediction:{}'.format(b_old_prediction))
history = self.actor.fit(x=[b_s, b_adv, b_old_prediction],
y=b_a_onehot,
epochs=5,
verbose=0)
# print('actor_loss_mean:{}'.format(history.history['loss']))
actor_loss_mean = np.mean(history.history['loss'])
self.critic.fit(x=b_s, y=b_vt, epochs=5,
verbose=0) # critic目标就是让td-error尽可能小
self.states, self.actions, self.rewards, self.states_, self.dones, self.v_by_trace = [], [], [], [], [], []
self.update_target_network()
return actor_loss_mean
def update_target_network(self):
self.actor_old.set_weights(self.target_update_alpha *
np.array(self.actor.get_weights()) +
(1 - self.target_update_alpha) *
np.array(self.actor_old.get_weights()))
def get_old_prediction(self, s):
s = np.reshape(s, (-1, self.n_features))
v = self.actor_old.predict([
s,
np.tile(self.dummy_advantage, (s.shape[0], 1)),
np.tile(self.dummy_old_prediction, (s.shape[0], 1))
])
return v
def get_v(self, s):
s = np.reshape(s, (-1, self.n_features))
v = np.squeeze(self.critic.predict(s))
return v
def cal_v_by_traceback(self):
'''
截断后或episode结束后，通过回溯计算V(s)=r+g*V(s_)
:return:
'''
# self.v_by_traceback = np.zeros_like(self.rewards)
if self.dones[-1]:
v = 0
else:
s = np.array(self.states_[-1])
v = self.get_v(s)
for t in reversed(range(0, len(self.rewards))):
v = v * self.gamma + self.rewards[t]
self.v_by_trace.append(v)
self.v_by_trace.reverse()
data = np.array(
pandas.read_csv(filepath_or_buffer="iris.data", header=None, nrows=150))
output = []
train_output = []
for i in range(150):
if (data[i, 4] == "Iris-setosa"):
output.append([1, 0, 0])
if (data[i, 4] == "Iris-versicolor"):
output.append([0, 1, 0])
if (data[i, 4] == "Iris-virginica"):
output.append([0, 0, 1])
train_input = np.concatenate(
(data[0:40, 0:4], data[50:90, 0:4], data[100:140, 0:4]), axis=0)
train_output = np.concatenate((output[0:40], output[50:90], output[100:140]),
axis=0)
test_input = np.concatenate(
(data[40:50, 0:4], data[90:100, 0:4], data[140:150, 0:4]), axis=0)
print(train_input.shape)
print(train_output.shape)
model.fit(x=train_input, y=train_output, epochs=10000)
model.save("model.hdf5")
# predict = model.predict(x=test_input)
#
# # Get the maximum values of each column i.e. along axis 0
# maxInColumns = np.amax(predict, axis=0)
# print('Max value of every column: ', maxInColumns)
# print(predict)
class NN:
"""
The biggest change for the duelling DQN is, that we must define a more complex DQN architecture
The architecture must define our Value and Advantage Layers
The way we define our model:
<<<>>>
The Keras functional API is a way to create models that are more flexible than the tf.keras.Sequential API.
The functional API can handle models with NON-LINEAR topology, SHARED layers, and even MULTIPLE inputs or outputs.
Read more here: https://keras.io/guides/functional_api/
<<<>>>
"""
def __init__(self, env, alpha: float = 0.001, decay: float = 0.0001):
"""
We initialize our functional model, therefore we need Input Shape and Output Shape
:param env:
:param alpha:
:param decay:
"""
self.alpha = alpha
self.decay = decay
self.model = None
# new to D-DDQN
self.init_model(env.observation_space.shape[0], env.action_space.n)
def init_model(self, input_shape: int, n_actions: int):
inp = Input(shape=(input_shape, ))
layer_shared1 = Dense(64, activation='relu')(inp)
layer_shared1 = BatchNormalization()(layer_shared1)
layer_shared2 = Dense(64, activation='relu')(layer_shared1)
layer_shared2 = BatchNormalization()(layer_shared2)
layer_v1 = Dense(64, activation='relu')(layer_shared2)
layer_v1 = BatchNormalization()(layer_v1)
layer_a1 = Dense(64, activation='relu')(layer_shared2)
layer_a1 = BatchNormalization()(layer_a1)
# the value layer ouput is a scalar value
layer_v2 = Dense(1, activation='linear')(layer_v1)
# The advantage function subtracts the value of the state from the Q
# function to obtain a relative measure of the importance of each action.
layer_a2 = Dense(n_actions, activation='linear')(layer_a1)
# the q layer combines the two streams of value and advantage function
# the lambda functional layer can perform lambda expressions on keras layers
# read more here : https://keras.io/api/layers/core_layers/lambda/
# the lambda equation is defined in https://arxiv.org/pdf/1511.06581.pdf on equation (9)
layer_q = Lambda(lambda x: x[0][:] + x[1][:] - K.mean(x[1][:]),
output_shape=(n_actions, ))([layer_v2, layer_a2])
self.model = Model(inp, layer_q)
self.model.compile(optimizer=Adam(lr=self.alpha), loss='mse')
def predict(self, *args, **kwargs):
"""
By wrapping the keras predict method we can handle our net as a standalone object
:param args: interface to keras.model.predict
:return: prediction
"""
return self.model.predict(*args, **kwargs)
def fit(self, *args, **kwargs):
"""
By wrapping the keras fit method we can handle our net as a standalone object
:param args: interface to keras.model.fit
:return: history object
"""
return self.model.fit(*args, **kwargs)
def get_weights(self):
"""
Passing the arguments to keras get_weights
"""
return self.model.get_weights()
def set_weights(self, *args, **kwargs):
"""
Passing the arguments to keras set_weights
"""
self.model.set_weights(*args, *kwargs)
class VBNChromosome:
""" Class that wraps the neural network. Includes functionality
for Virtual Batch Normalization and the mutation of weights."""
def __init__(self, number_actions=6, input_channels=4):
self.input_channels = input_channels
self.number_actions = number_actions
inputs, outputs = self.construct_layers()
self.model = Model(inputs=inputs, outputs=outputs)
def construct_layers(self):
""" Construct the Mnih et al. DQN architecture."""
inputs = Input(shape=(84, 84, self.input_channels))
layer1 = Conv2D(32, [8, 8], strides=(4, 4), activation="relu")(inputs)
layer1 = BatchNormalization(momentum=0.95, center=False, scale=False)(layer1)
layer2 = Conv2D(64, [4, 4], strides=(2, 2), activation="relu")(layer1)
layer2 = BatchNormalization(momentum=0.95, center=False, scale=False)(layer2)
layer3 = Conv2D(64, [3, 3], strides=(1, 1), activation="relu")(layer2)
layer3 = BatchNormalization(momentum=0.95, center=False, scale=False)(layer3)
layer4 = Flatten()(layer3)
layer5 = Dense(512, activation="relu")(layer4)
layer5 = BatchNormalization(momentum=0.95, center=False, scale=False)(layer5)
action = Dense(self.number_actions, activation="softmax")(layer5)
return [inputs], action
def virtual_batch_norm(self, samples):
""" We apply Batch Normalization on a number of samples. By setting the learning
rate to 0 we make sure that the weights and biases are not affected. This method
is only ment to be used at the start of training."""
optimizer = tf.keras.optimizers.SGD(learning_rate=0)
loss = tf.keras.losses.MeanSquaredError()
self.model.compile(loss=loss, optimizer=optimizer)
fake_y = np.zeros((len(samples), self.number_actions))
self.model.fit(np.array(samples), fake_y)
def get_weights(self, layers=None):
""" Retrieve all the weights of the network. """
layers = layers if layers else self.model.layers
layer_weights = chain(*[layer.get_weights() for layer in layers])
flat_weights = [weights.flatten() for weights in layer_weights]
return np.concatenate(flat_weights)
def set_weights(self, flat_weights, layers=None):
""" Set all the weights of the network. """
i = 0
layers = layers if layers else self.model.layers
for layer in layers:
new_weights = []
for sub_layer in layer.get_weights():
reshaped = flat_weights[i: i + sub_layer.size].reshape(sub_layer.shape)
new_weights.append(reshaped)
i += sub_layer.size
layer.set_weights(new_weights)
def get_perturbable_layers(self):
""" Get all the perturbable layers of the network. This excludes the
BatchNorm layers. """
return [layer for layer in self.model.layers if
not isinstance(layer, BatchNormalization)]
def get_perturbable_weights(self):
""" Get all the perturbable weights of the network. This excludes the
BatchNorm weights. """
return self.get_weights(self.get_perturbable_layers())
def set_perturbable_weights(self, flat_weights):
""" Set all the perturbable weights of the network. This excludes setting
the BatchNorm weights. """
self.set_weights(flat_weights, self.get_perturbable_layers())
def mutate(self, mutation_power):
""" Mutate the current weights by adding a normally distributed vector of
noise to the current weights. """
weights = self.get_perturbable_weights()
noise = np.random.normal(loc=0.0, scale=mutation_power, size=weights.shape)
self.set_perturbable_weights(weights + noise)
return noise
def determine_actions(self, inputs):
""" Choose an action based on the pixel inputs. We do this by simply
selecting the action with the highest outputted value. """
actions = self.model(inputs)
return [np.argmax(action_set) for action_set in actions]
patience=5,
verbose=True)
reduce_lr = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss',
factor=0.2,
patience=3,
min_lr=lr,
verbose=True)
file_name = 'models/weights-improvement-{epoch:02d}-{val_loss:.2f}.hdf5'
save_model = tf.keras.callbacks.ModelCheckpoint('{}'.format(file_name),
monitor='val_loss')
log_dir = "logs/fit/" + dt.datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard = tf.keras.callbacks.TensorBoard(log_dir=log_dir,
histogram_freq=1)
model_resnet.fit(
ds_img_train,
epochs=epochs,
callbacks=[reduce_lr, early_stopping, save_model, tensorboard],
validation_data=ds_img_valid,
validation_steps=len(train_dg) // batch_size)
# print('DenseNet121') # runs with batch size of <32
# base_model = DenseNet121(weights='imagenet', include_top=False, input_shape=input_shape)
# x2 = Flatten()(base_model.get_output_at(-1))
# x2 = Dense(32, activation='relu')(x2)
# output2 = Dense(lab_dim, activation='sigmoid')(x2)
# model_resnet = Model(base_model.input, output2)
# model_resnet.compile(optimizer=tf.optimizers.Adam(learning_rate=lr),
# loss=tf.nn.sigmoid_cross_entropy_with_logits, metrics=['accuracy'])
# early_stopping = tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=3)
# reduce_lr = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=3, min_lr=lr)
# file_name = 'weights-improvement-{epoch:02d}-{val_loss:.2f}.hdf5'
# save_model = tf.keras.callbacks.ModelCheckpoint('{}'.format(file_name), monitor='val_loss')
question_enc = np.array([
x for x in df['question'].map(
lambda x: encode_sentence(x, max_question)).values
])
answer_enc = np.array([
x
for x in df['answer'].map(lambda x: encode_sentence(x, max_answer)).values
])
print(df)
a = Input(shape=(max_question, WORD_SIZE))
b = Dense(1024)(a)
b = Dense(2048)(b)
b = Dense(4096)(b)
c = Dense(WORD_SIZE)(b)
model = Model(inputs=a, outputs=c)
model.compile(loss='mean_squared_error', optimizer='sgd')
model.fit(question_enc, answer_enc, epochs=10)
results = model.predict(question_enc)
def get_words_from_vecs(x):
return [get_word(vec) for vec in x]
print("\n".join([" ".join(get_words_from_vecs(x)) for x in results]))
y_test = df2['label'].values
inputt = Input(shape=(21, ))
x = Dense(units=20, activation='sigmoid')(inputt)
x = Dense(units=18, activation='sigmoid')(x)
x = Dense(units=14, activation='sigmoid')(x)
x = Dense(units=5, activation='softmax')(x)
model = Model(inputs=inputt, outputs=x)
Optimizer = SGD(lr=0.01)
model.compile(optimizer=Optimizer,
loss='mean_squared_error',
metrics=['accuracy'])
print(X_train.shape)
print(y_train.shape)
History = model.fit(x=X_train,
y=y_train,
epochs=500,
validation_split=0.1,
shuffle=True,
batch_size=512)
print(History)
y_pred = model.predict(testData)
y_pred_bool = np.argmax(y_pred, axis=1)
print(classification_report(y_test, y_pred_bool))
def run_single_test(algorithm_def,
gen_train,
gen_val,
load_weights,
freeze_weights,
x_test,
y_test,
lr,
batch_size,
epochs,
epochs_warmup,
model_checkpoint,
scores,
loss,
metrics,
logging_path,
kwargs,
clipnorm=None,
clipvalue=None,
model_callback=None):
print(metrics)
print(loss)
metrics = make_custom_metrics(metrics)
loss = make_custom_loss(loss)
if load_weights:
enc_model = algorithm_def.get_finetuning_model(model_checkpoint)
else:
enc_model = algorithm_def.get_finetuning_model()
pred_model = apply_prediction_model(
input_shape=enc_model.outputs[0].shape[1:],
algorithm_instance=algorithm_def,
**kwargs)
outputs = pred_model(enc_model.outputs)
model = Model(inputs=enc_model.inputs[0], outputs=outputs)
print_flat_summary(model)
if epochs > 0:
callbacks = [TerminateOnNaN()]
logging_csv = False
if logging_path is not None:
logging_csv = True
logging_path.parent.mkdir(exist_ok=True, parents=True)
logger_normal = CSVLogger(str(logging_path), append=False)
logger_after_warmup = LogCSVWithStart(
str(logging_path), start_from_epoch=epochs_warmup, append=True)
if freeze_weights or load_weights:
enc_model.trainable = False
if freeze_weights:
print(("-" * 10) +
"LOADING weights, encoder model is completely frozen")
if logging_csv:
callbacks.append(logger_normal)
elif load_weights:
assert epochs_warmup < epochs, "warmup epochs must be smaller than epochs"
print(("-" * 10) +
"LOADING weights, encoder model is trainable after warm-up")
print(("-" * 5) + " encoder model is frozen")
w_callbacks = list(callbacks)
if logging_csv:
w_callbacks.append(logger_normal)
model.compile(optimizer=get_optimizer(clipnorm, clipvalue, lr),
loss=loss,
metrics=metrics)
model.fit(
x=gen_train,
validation_data=gen_val,
epochs=epochs_warmup,
callbacks=w_callbacks,
)
epochs = epochs - epochs_warmup
enc_model.trainable = True
print(("-" * 5) + " encoder model unfrozen")
if logging_csv:
callbacks.append(logger_after_warmup)
else:
print(("-" * 10) +
"RANDOM weights, encoder model is fully trainable")
if logging_csv:
callbacks.append(logger_normal)
# recompile model
model.compile(optimizer=get_optimizer(clipnorm, clipvalue, lr),
loss=loss,
metrics=metrics)
model.fit(x=gen_train,
validation_data=gen_val,
epochs=epochs,
callbacks=callbacks)
model.compile(optimizer=get_optimizer(clipnorm, clipvalue, lr),
loss=loss,
metrics=metrics)
y_pred = model.predict(x_test, batch_size=batch_size)
scores_f = make_scores(y_test, y_pred, scores)
if model_callback:
model_callback(model)
# cleanup
del pred_model
del enc_model
del model
algorithm_def.purge()
K.clear_session()
for i in range(15):
gc.collect()
for s in scores_f:
print("{} score: {}".format(s[0], s[1]))
return scores_f
save_best_only=True)
lr_scheduler = LearningRateScheduler(lr_schedule)
lr_reducer = ReduceLROnPlateau(factor=np.sqrt(0.1),
cooldown=0,
patience=5,
min_lr=0.5e-6)
callbacks = [checkpoint, lr_reducer, lr_scheduler]
# choose training configs
sgd = SGD(lr=0.001, momentum=0.9)
model.compile(optimizer=sgd,
loss=categorical_crossentropy,
metrics=['accuracy'])
model.summary()
# train
hist = model.fit(x_train,
y_train,
batch_size=100,
epochs=10,
shuffle=True,
verbose=1,
validation_split=0.1,
callbacks=callbacks)
# test
model.evaluate(x_test, y_test, verbose=1)
def main(arg):
directory = Path('./saved_predictions/')
directory.mkdir(exist_ok=True)
directory = Path('./saved_models/')
directory.mkdir(exist_ok=True)
directory = Path('./training_checkpoints/')
directory.mkdir(exist_ok=True)
input_yx_size = tuple(args.input_yx_size)
batch_size = args.batch_size
epochs = args.epochs
learning_rate = args.learning_rate
num_test_samples = args.num_test_samples
save_weights = args.save_weights
every = args.every
num_samples = args.num_samples
save_train_prediction = args.save_train_prediction
save_test_prediction = args.save_test_prediction
verbose = args.verbose
validation_ratio = args.validation_ratio
y_axis_len, x_axis_len = input_yx_size
decay = args.decay
decay = args.decay
load_weights = args.load_weights
y_axis_len, x_axis_len = input_yx_size
num_points = y_axis_len * x_axis_len
is_flat_channel_in = args.is_flat_channel_in
input_points = Input(shape=(num_points, 4))
x = input_points
x = Convolution1D(64, 1, activation='relu', input_shape=(num_points, 4))(x)
x = BatchNormalization()(x)
x = Convolution1D(128, 1, activation='relu')(x)
x = BatchNormalization()(x)
x = Convolution1D(512, 1, activation='relu')(x)
x = BatchNormalization()(x)
x = MaxPooling1D(pool_size=num_points)(x)
x = Dense(512, activation='relu')(x)
x = BatchNormalization()(x)
x = Dense(256, activation='relu')(x)
x = BatchNormalization()(x)
x = Dense(16,
weights=[
np.zeros([256, 16]),
np.array([1, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0,
1]).astype(np.float32)
])(x)
input_T = Reshape((4, 4))(x)
# forward net
g = Lambda(mat_mul, arguments={'B': input_T})(input_points)
g = Convolution1D(64, 1, input_shape=(num_points, 3), activation='relu')(g)
g = BatchNormalization()(g)
g = Convolution1D(64, 1, input_shape=(num_points, 3), activation='relu')(g)
g = BatchNormalization()(g)
# feature transformation net
f = Convolution1D(64, 1, activation='relu')(g)
f = BatchNormalization()(f)
f = Convolution1D(128, 1, activation='relu')(f)
f = BatchNormalization()(f)
f = Convolution1D(128, 1, activation='relu')(f)
f = BatchNormalization()(f)
f = MaxPooling1D(pool_size=num_points)(f)
f = Dense(512, activation='relu')(f)
f = BatchNormalization()(f)
f = Dense(256, activation='relu')(f)
f = BatchNormalization()(f)
f = Dense(64 * 64,
weights=[
np.zeros([256, 64 * 64]),
np.eye(64).flatten().astype(np.float32)
])(f)
feature_T = Reshape((64, 64))(f)
# forward net
g = Lambda(mat_mul, arguments={'B': feature_T})(g)
seg_part1 = g
g = Convolution1D(64, 1, activation='relu')(g)
g = BatchNormalization()(g)
g = Convolution1D(32, 1, activation='relu')(g)
g = BatchNormalization()(g)
g = Convolution1D(32, 1, activation='relu')(g)
g = BatchNormalization()(g)
# global_feature
global_feature = MaxPooling1D(pool_size=num_points)(g)
global_feature = Lambda(exp_dim, arguments={'num_points':
num_points})(global_feature)
# point_net_seg
c = concatenate([seg_part1, global_feature])
""" c = Convolution1D(512, 1, activation='relu')(c)
c = BatchNormalization()(c)
c = Convolution1D(256, 1, activation='relu')(c)
c = BatchNormalization()(c)
c = Convolution1D(128, 1, activation='relu')(c)
c = BatchNormalization()(c)
c = Convolution1D(128, 1, activation='relu')(c)
c = BatchNormalization()(c) """
c = Convolution1D(256, 1, activation='relu')(c)
c = BatchNormalization()(c)
c = Convolution1D(128, 4, activation='relu', strides=4)(c)
c = BatchNormalization()(c)
c = Convolution1D(128, 4, activation='relu', strides=4)(c)
c = BatchNormalization()(c)
c = Convolution1D(128, 4, activation='relu', strides=4)(c)
c = BatchNormalization()(c)
c = Convolution1D(64, 4, activation='relu', strides=4)(c)
c = BatchNormalization()(c)
c = Convolution1D(64, 4, activation='relu', strides=4)(c)
c = BatchNormalization()(c)
c = Convolution1D(32, 1, activation='relu')(c)
c = BatchNormalization()(c)
""" c = Convolution1D(128, 4, activation='relu',strides=4)(c)
c = Convolution1D(64, 4, activation='relu',strides=4)(c)
c = Convolution1D(32, 4, activation='relu',strides=4)(c)
c = Convolution1D(16, 1, activation='relu')(c)
c = Convolution1D(1, 1, activation='relu')(c) """
#c = tf.keras.backend.squeeze(c,3);
c = CuDNNLSTM(64, return_sequences=False)(c)
#c =CuDNNLSTM(784, return_sequences=False))
#c =CuDNNLSTM(256, return_sequences=False))
#c = Reshape([16,16,1])(c)
c = Reshape([8, 8, 1])(c)
c = Conv2DTranspose(8, (3, 3),
padding="same",
activation="relu",
strides=(2, 2))(c)
c = Conv2DTranspose(8, (3, 3), padding="valid", activation="relu")(c)
#c =Dropout(0.4))
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(16, (3, 3), padding="valid", activation="relu")(c)
#c =Dropout(0.4))
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(32, (3, 3), padding="valid", activation="relu")(c)
#c =Dropout(0.4))
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(32, (3, 3), padding="valid", activation="relu")(c)
#c =Dropout(0.4))
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(32, (3, 3), padding="valid", activation="relu")(c)
#c =Dropout(0.4))
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(64, (3, 3), padding="valid", activation="relu")(c)
#c =Dropout(0.4))
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(64, (3, 3), padding="valid", activation="relu")(c)
#c =Dropout(0.4))
c = tf.keras.layers.BatchNormalization()(c)
#c =Dropout(0.4))
c = Conv2DTranspose(128, (3, 3),
padding="same",
activation="relu",
strides=(2, 2))(c)
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(128, (3, 3), padding="valid", activation="relu")(c)
#c =Dropout(0.4))
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(128, (3, 3),
padding="same",
activation="relu",
strides=(2, 2))(c)
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(128, (3, 3), padding="valid", activation="relu")(c)
c = tf.keras.layers.BatchNormalization()(c)
#c =Dropout(0.4))
#c =tf.keras.layers.BatchNormalization())
c = Conv2DTranspose(64, (3, 3), padding="same", strides=(4, 2))(c)
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(32, (3, 3), padding="valid", activation="relu")(c)
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(32, (3, 3), padding="valid", activation="relu")(c)
c = tf.keras.layers.BatchNormalization()(c)
#c =Dropout(0.4))
c = Conv2DTranspose(32, (3, 3),
padding="same",
activation="relu",
strides=(1, 1))(c)
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(32, (3, 1), padding="valid", activation="relu")(c)
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(32, (3, 1), padding="valid", activation="relu")(c)
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(16, (1, 1), padding="valid", activation="relu")(c)
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(8, (1, 1), padding="valid", activation="relu")(c)
c = tf.keras.layers.BatchNormalization()(c)
c = Conv2DTranspose(1, (1, 1), padding="valid")(c)
""" c =Conv2DTranspose(4, (1,1),padding="same",activation="relu"))
c =Conv2DTranspose(2, (1,1),padding="same",activation="relu"))
#c =Dropout(0.4))
c =Conv2DTranspose(1, (1,1),padding="same")) """
prediction = tf.keras.layers.Reshape([512, 256])(c)
""" c1 ,c2 = tf.split(c,[256,256],axis=1,name="split")
complexNum = tf.dtypes.complex(
c1,
c2,
name=None
)
complexNum =tf.signal.ifft2d(
complexNum,
name="IFFT"
)
real = tf.math.real(complexNum)
imag = tf.math.imag(complexNum)
con = concatenate([real,imag])
prediction =tf.keras.layers.Reshape([ 512, 256])(con)
"""
# define model
model = Model(inputs=input_points, outputs=prediction)
opt = tf.keras.optimizers.Adam(lr=learning_rate, decay=decay)
loss = tf.keras.losses.MeanSquaredError()
mertric = ['mse']
if args.loss is "MAE":
loss = tf.keras.losses.MeanAbsoluteError()
mertric = ['mae']
model.compile(
loss=loss,
optimizer=opt,
metrics=mertric,
)
model.summary()
if load_weights:
model.load_weights('./training_checkpoints/cp-best_loss.ckpt')
#edit data_loader.py if you want to play with data
input_ks, ground_truth = load_data(num_samples,
is_flat_channel_in=is_flat_channel_in)
input_ks = input_ks / np.max(input_ks)
checkpoint_path = "./training_checkpoints/cp-{epoch:04d}.ckpt"
checkpoint_dir = os.path.dirname(checkpoint_path)
# Create checkpoint callback
#do you want to save the model's wieghts? if so set this varaible to true
cp_callback = []
NAME = "NUFFT_NET"
tensorboard = TensorBoard(log_dir="logs/{}".format(NAME))
cp_callback.append(tensorboard)
if save_weights:
cp_callback.append(
tf.keras.callbacks.ModelCheckpoint(checkpoint_dir,
save_weights_only=True,
verbose=verbose,
period=every))
if args.is_train:
model.fit(input_ks,
ground_truth,
batch_size=batch_size,
epochs=epochs,
validation_split=validation_ratio,
callbacks=cp_callback)
if args.name_model is not "":
model.save('./saved_mdoels/' + args.name_model)
dict_name = './saved_predictions/'
#return to image size
x_axis_len = int(x_axis_len / 4)
np.random.seed(int(time()))
if save_train_prediction <= num_samples:
rand_ix = np.random.randint(0, num_samples - 1, save_train_prediction)
#kspace = np.zeros((save_train_prediction,
#y_axis_len,input_ks[rand_ix].shape[1]))
kspace = input_ks[rand_ix]
if args.save_input:
np.save("./saved_predictions/inputs.npy", input_ks[rand_ix])
ground_truth = ground_truth[rand_ix]
preds = model.predict(kspace, batch_size=save_train_prediction)
for i in range(save_train_prediction):
output = np.reshape(preds[i], (y_axis_len * 2, x_axis_len))
output = output * 255
output[np.newaxis, ...]
output_gt = ground_truth[i]
output_gt[np.newaxis, ...]
output = np.concatenate([output, output_gt], axis=0)
np.save(dict_name + 'prediction%d.npy' % (i + 1), output)
input_ks, ground_truth = load_data(
num_test_samples, 'test', is_flat_channel_in=is_flat_channel_in)
input_ks = input_ks / np.max(input_ks)
if args.is_eval:
model.evaluate(input_ks,
ground_truth,
batch_size,
verbose,
callbacks=cp_callback)
if save_test_prediction <= num_test_samples:
rand_ix = np.random.randint(0, num_test_samples - 1,
save_test_prediction)
kspace = input_ks[rand_ix]
if args.save_input:
np.save("./saved_predictions/test_inputs.npy", input_ks[rand_ix])
ground_truth = ground_truth[rand_ix]
preds = model.predict(kspace, batch_size=save_test_prediction)
for i in range(save_test_prediction):
output = np.reshape(preds[i], (y_axis_len * 2, x_axis_len))
output = output * 255
output[np.newaxis, ...]
output_gt = ground_truth[i]
output_gt[np.newaxis, ...]
output = np.concatenate([output, output_gt], axis=0)
np.save(dict_name + 'test_prediction%d.npy' % (i + 1), output)
def knowledge_transfer(current_student: Model, method: Method, loss: Union[LossType, List[LossType]]) -> \
Tuple[Model, History]:
"""
Performs KT.
:param current_student: the student to be used for the current KT method.
:param method: the method to be used for the KT.
:param loss: the KT loss to be used.
:return: Tuple containing a student Keras model and its training History object.
"""
kt_logging.debug('Configuring student...')
weights = None
y_train_adapted = y_train_concat
y_val_adapted = y_val_concat
metrics = {}
if method == Method.DISTILLATION:
# Adapt student
current_student = kd_student_adaptation(current_student, temperature)
# Create KT metrics.
metrics = generate_supervised_metrics(method)
monitoring_metric = 'val_accuracy'
elif method == Method.PKT_PLUS_DISTILLATION:
# Adapt student
current_student = pkt_plus_kd_student_adaptation(current_student, temperature)
# Create importance weights for the different losses.
weights = [kd_importance_weight, pkt_importance_weight]
if selective_learning:
selective_learning_weights = []
for _ in range(n_submodels):
selective_learning_weights.extend(weights)
weights = selective_learning_weights
# Adapt the labels.
y_train_adapted.extend(y_train_adapted)
y_val_adapted.extend(y_val_adapted)
else:
# Adapt the labels.
y_train_adapted = [y_train_concat, y_train_concat]
y_val_adapted = [y_val_concat, y_val_concat]
# Create KT metrics.
metrics = generate_supervised_metrics(method)
monitoring_metric = 'val_concatenate_accuracy'
else:
# PKT performs KT, but also rotates the space, thus evaluating results has no meaning,
# since the neurons representing the classes are not the same anymore.
monitoring_metric = 'val_loss'
if selective_learning:
current_student = selective_learning_student_adaptation(current_student, n_submodels)
monitoring_metric = 'val_loss'
# Create optimizer.
optimizer = initialize_optimizer(optimizer_name, learning_rate, decay, beta1, beta2, rho, momentum,
clip_norm, clip_value)
# Compile student.
current_student.compile(optimizer, loss, metrics, weights)
# Initialize callbacks list.
kt_logging.debug('Initializing Callbacks...')
# Create a temp file, in order to save the model, if needed.
tmp_weights_path = None
if use_best_model:
tmp_weights_path = join(gettempdir(), next(mktemp()) + '.h5')
callbacks_list = init_callbacks(monitoring_metric, lr_patience, lr_decay, lr_min, early_stopping_patience,
verbosity, tmp_weights_path, selective_learning)
# Train student.
history = current_student.fit(x_train, y_train_adapted, batch_size=batch_size, callbacks=callbacks_list,
epochs=epochs, validation_data=(x_val, y_val_adapted), verbose=verbosity)
if exists(tmp_weights_path):
# Load best weights and delete the temp file.
current_student.load_weights(tmp_weights_path)
remove(tmp_weights_path)
# Rewind student to its normal state, if necessary.
if selective_learning:
current_student = selective_learning_student_rewind(current_student, optimizer=optimizer, loss=loss[0],
metrics=metrics)
if method == Method.DISTILLATION:
current_student = kd_student_rewind(current_student)
elif method == Method.PKT_PLUS_DISTILLATION:
current_student = pkt_plus_kd_rewind(current_student)
return current_student, history
internal = self.FC_1(attention_vector)
# internal = self.FC_2(internal)
final_output = self.classification_layer(internal)
return final_output
# create the model
recurrent_fusion_model = Model()
recurrent_fusion_model.compile(optimizer=keras.optimizers.Adam(lr=lr),
loss=sparse_categorical_cross_entropy_loss,
metrics=[acc_top_1, acc_top_5])
# build internal tensors
recurrent_fusion_model.fit(*next(train_generator()),
batch_size=1,
epochs=1,
verbose=0)
# get tensorflow saver ready > will be used if a checkpoint found on drive
saver = tf.train.Saver(recurrent_fusion_model.variables)
if checkpoint_found:
# restore the model from the checkpoint
log("Model restored")
eval_globals.best_video_level_accuracy_1 = float(
zip_file_name.split("-")[1])
log("Current Best", eval_globals.best_video_level_accuracy_1)
saver.restore(tf.keras.backend.get_session(),
checkpoints) # use tensorflow saver
initial_epoch = int(zip_file_name.split("-")[0]) # get epoch number
y = tf.keras.layers.Dense(units=32,
activation='elu',
kernel_initializer='he_uniform')(y)
y = tf.keras.layers.Dense(units=2,
activation='softmax',
kernel_initializer='he_uniform')(y)
wenz_model = Model(inputs=[input1, input2], outputs=y)
adam = Adam(lr=0.02, decay=0.01)
wenz_model.compile(optimizer='adam',
loss=tf.keras.losses.BinaryCrossentropy(),
metrics=['accuracy'])
checkpoint_path = "training/cp.ckpt"
checkpoint_dir = os.path.dirname(checkpoint_path)
# Create a callback that saves the model's weights
cp_callback = tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_path,
save_weights_only=True,
verbose=1)
wenz_model.fit(XY, epochs=10,
callbacks=[cp_callback]) # add validation training_data?
test_loss, test_acc = wenz_model.evaluate(XYt, verbose=2)
print('\nTest accuracy:', test_acc)
wenz_model.save(
'/home/pirate/PycharmProjects/SchafkopfAI/models/trained_models/test-wenz-prediction6'
)
store.close()
x2 = []
for i in x:
x1.append(i[0])
x2.append(i[1])
y = np.array([7.8, 8.6, 8.7, 7.9, 8.4, 8.9, 10.4, 11.6, 13.9, 15.8])
print(x.shape)
print(y.shape)
with tf.device('/cpu:0'):
inputs = layers.Input(shape=(2, ))
out = layers.Dense(1,
use_bias=False,
kernel_initializer=initializers.RandomUniform())(inputs)
model = Model(inputs=inputs, outputs=out)
dataset = tf.data.Dataset.from_tensor_slices((x, y)).batch(3).repeat()
# print(dataset.take(1))
model.compile(
optimizer=tf.train.GradientDescentOptimizer(learning_rate=0.0001),
loss='mean_squared_error')
print_weights = callbacks.LambdaCallback(
on_epoch_end=lambda batch, logs: print(model.layers[1].get_weights()))
model.summary()
model.fit(dataset, epochs=20, steps_per_epoch=1, callbacks=[print_weights])
# model.fit(dataset,epochs=20,steps_per_epoch=1)
shared_model(right_input)])
model = Model(inputs=[left_input, right_input], outputs=[malstm_distance])
if gpus >= 2:
# `multi_gpu_model()` is a so quite buggy. it breaks the saved model.
model = tf.keras.utils.multi_gpu_model(model, gpus=gpus)
model.compile(loss='mean_squared_error',
optimizer=tf.keras.optimizers.Adam(),
metrics=['accuracy'])
model.summary()
shared_model.summary()
# Start trainings
training_start_time = time()
malstm_trained = model.fit([X_train['left'], X_train['right']],
Y_train,
batch_size=batch_size,
epochs=n_epoch)
training_end_time = time()
logging.info("Training time finished.\n%d epochs in %12.2f" %
(n_epoch, training_end_time - training_start_time))
saver = tf.compat.v1.train.Saver()
session = tf.compat.v1.keras.backend.get_session()
saver.save(session, SESSION_PATH)
model.save(get_model_path(n_epoch, embedding_dim, max_seq_length,
n_hidden))
# plot(malstm_trained)